Nonlinear Reduced-Order Modeling of Automotive Aeroacoustics Using β-Variational Autoencoders
Please login to view abstract download link
Unsteady aerodynamic noise, known as wind noise, poses significant design challenges for electric vehicles,where the masking noise sources related to the engine are not present and wind noise becomes dominant at lower speeds. Side-view mirrors generate strong unsteady vortices and broadband noise through turbulent flow separation, requiring high-fidelity simulations that are computationally prohibitive for design optimization loops. This work develops efficient reduced-order models (ROMs) for aeroacoustic prediction around a generic side-view mirror at Reynolds number ReD = 5.2 × 105. Flow and acoustic fields are obtained using the Lattice Boltzmann solver PowerFLOW with Very Large Eddy Simulation (VLES), validated against reference numerical studies [1]. A dataset of 12,500 snapshots covering 600 convective time units provides the foundation for ROMdevelopment. Two approaches are investigated: Proper Orthogonal Decomposition (POD), establishing a linear baseline, and a convolutional β-Variational Autoencoder (β-VAE) [2], learning nonlinear manifolds of the wall-pressure field. Results show the limitations of linear projection methods for turbulent flows. POD requires over 100 modes to capture 80% of acoustic energy and more than 500 modes for 90% reconstruction. The β-VAE, in contrast, achieves 87% energy reconstruction with only 16 latent variables, capturing fine-scale structures through nonlinear dimensionality reduction. This work further integrates transformers to predict the temporal dynamics of latent variables and employs a joint training strategy combining large-scale flow features with higher-frequency content to capture acoustic propagation, enabling construction of a complete reduced-order model for aeroacoustics.
